Multimodal Deep Learning‐Assisted Microstructural Evaluation System for Sintered NdFeB Magnets

ABSTRACT

The performance of sintered NdFeB magnets, especially coercivity, is critically determined by their complex microstructure. However, a quantitative microstructural evaluation system (MES) is much needed but has long been missing in the community. Here, we developed a multimodal deep learning workflow to discover the critical descriptors of grains with poor demagnetization resistance, thereby establishing the MES. Specifically, to extract precise microstructural descriptors from images, we developed an image segmentation model called magnets segment anything model (MagSAM), achieving superior pixel‐level segmentation accuracy for Nd ₂ Fe ₁₄ B grains. Based on nucleation theory, four knowledge‐informed descriptors (i.e., misorientation angle, coating completeness, corner curvature radius, and grain boundary straightness) were defined to reflect the demagnetization resistance of the grain. By applying MagSAM to correlative in situ backscattered electron (BSE), electron backscatter diffraction (EBSD), and dynamic magneto‐optic Kerr effect microscopy (MOKE) images, we extracted the microstructural descriptors and demagnetization resistance of individual grains to construct a multimodal dataset. The established MES revealed that the averaged Feret diameter is the most suitable metric for grain size, as it dominates demagnetization resistance. Corner curvature radius and coating completeness show nearly identical importance. Grains that are larger, with lower coating completeness and corner curvature radius but higher grain boundary straightness values, tend to exhibit poor demagnetization resistance, thereby deteriorating coercivity. Based on this, the developed MES can assess the coercivity level directly from microstructural images. We believe that the MES will provide a standardized diagnostic tool to quantitatively explain the coercivity variations caused by microstructure.